Curator-authored starting points for practical AI engineering. These are research hypotheses to investigate, not validated businesses or requests from upstream maintainers.
Reproduce a baseline. Correct a source record. Write a clearer setup guide. Add reviewed language examples. Choose work that leaves something another person can verify.
Opportunity priority reflects task clarity, data accessibility, evaluation feasibility and Indian-market relevance. The score helps order curator attention; it is not a market-size estimate or a promise of commercial demand.
RESEARCHED STARTING POINTS
The Opportunity Radar
CURATOR PROPOSALcurator proposed
Small models, safer multilingual routing
A support request can be syntactically valid yet route to the wrong tool. Build a compact evaluation set for route selection, ambiguity and abstention across English, Hindi and mixed-language input.
Who it helps
Developers building support triage and internal workflow assistants.
Indian-market relevance
Hindi in Devanagari and romanized input offer concrete starting cases; expansion to other languages requires reviewed examples.
Technical difficulty
moderate
Available data
Create original, consent-free synthetic requests; have fluent reviewers validate labels. Public model access does not supply a representative customer test set.
Evaluation feasibility
Deterministic JSON/schema and route checks; report per-language errors, abstention and unsupported actions.
Operating complexity
low: local model run, fixed dataset and deterministic scorer; no autonomous tool execution.
Commercial hypothesis: Hypothesis: organizations with repeatable triage workflows may value a scoped routing audit. Customer interviews are required.
A contribution path: Propose 20 clearly labeled cases and adversarial near-misses with reviewer notes; use the submission workflow to request review.
Generated SQL can look plausible while returning the wrong rows or attempting a write. Build a read-only query evaluation using synthetic schemas and expected result sets.
Who it helps
Engineers adding natural-language query helpers to internal data tools.
Indian-market relevance
Useful for small teams with operational spreadsheets or databases; Indian-market relevance is a hypothesis to validate through user research.
Technical difficulty
moderate
Available data
Original synthetic schemas and records avoid customer-data rights and privacy issues. Keep evaluation queries separate from development examples.
Evaluation feasibility
Compare result sets and permission violations in an isolated read-only database; use timeouts and query limits.
Operating complexity
low for offline synthetic tests; production integration requires a separate access-control design.
Commercial hypothesis: Hypothesis: data teams may value a bounded accuracy and safety assessment. Do not promise autonomous access to production systems.
A contribution path: Submit three synthetic schemas, representative questions, expected outputs and unsafe-query cases.
Keep the evidence attached to every extracted field
Document pipelines often lose the relationship between an extracted value and its source location. Prototype field extraction that retains page references and rejects unsupported values.
Who it helps
Small organizations processing their own forms, invoices or operational documents.
Indian-market relevance
Multiscript documents and varied scan quality create test cases worth examining; their frequency and commercial value must be validated.
Technical difficulty
moderate
Available data
Use original synthetic documents or explicitly permitted public samples. Avoid customer records and personal information in public tests.
Evaluation feasibility
Field exact match, missing-value behavior and evidence-page correctness; measure retrieval separately from extraction.
Operating complexity
medium: document conversion and indexing can run locally, but OCR and layout variation need testing.
Commercial hypothesis: Hypothesis: bounded document workflows may benefit from an auditable private pipeline. A discovery engagement must establish volume and error cost.
A contribution path: Contribute a permitted document set, expected field values and page references with a reproducible conversion command.
General translation quality can hide incorrect product names, units and domain terms. Design a small English-to-Indic terminology-preservation test with explicit acceptability rules.
Who it helps
Teams localizing product help, service instructions and internal documentation.
Indian-market relevance
Begin with one language pair and reviewed terms, then expand separately; do not generalize from Hindi to all Indic languages.
Technical difficulty
moderate
Available data
IN22 provides a possible evaluation reference under its stated terms and gated access; domain terminology examples need separate rights and expert review.
Evaluation feasibility
Exact preservation checks for names and units plus blinded human review of meaning. Report both, rather than one combined leaderboard score.
Operating complexity
medium: language reviewers and model access are more substantial than the deterministic harness.
Commercial hypothesis: Hypothesis: localization teams may pay for repeatable terminology checks and integration. No demand has been validated.
A contribution path: Submit a glossary you have rights to share, 10 test sentences and a review rubric; never upload confidential manuals.
Make speech errors visible by language and condition
A single speech-recognition score can hide failures on names, mixed-language speech or noisy recordings. Build an error-analysis workflow for a narrowly defined, consented audio sample.
Who it helps
Developers evaluating transcription for internal service workflows.
Indian-market relevance
Choose a specific Indian language and recording setting; dialect and code-switching need explicit sample descriptions.
Technical difficulty
high
Available data
Audio needs informed consent or a clear reuse license, verified transcripts and speaker privacy controls. No customer audio is included.
Evaluation feasibility
Word or character error rate with a documented normalization policy; inspect names and numbers separately.
Operating complexity
medium: audio storage, reviewer time and model runtime need budgeting; no always-on GPU service is required for an offline study.
Commercial hypothesis: Hypothesis: transcription audits may help teams identify whether automation is appropriate. Benefit is unproven until tested on representative audio.
A contribution path: Propose the dataset consent/licensing plan and transcription rubric before gathering recordings.